piecewise-linear-functions

This paper improves theoretical bounds on the depth of ReLU networks needed to represent the maximum function, showing exact two-hidden-layer representations for up to 10 inputs and improved depth for larger n via exact linear algebra techniques.